Research On Speech Enhancement Method Based On Fiber Ring Laser Microphone

With the development of optical fiber communication technology,tunable fiber lasers have been widely used in optical fiber communication,sensing technology and medical fields due to their tunable,narrow line width,good beam quality,good compatibility and low cost.In this thesis,an optical fiber microphone is designed based on the tunable laser ring cavity,and the speech enhancement method is studied according to the voice characteristics of the optical fiber microphone.The main contents of this article are as follows:First,the background and significance of the research on this subject are briefly described,and the fiber-optic microphone based on laser ring cavity designed in this thesis is introduced in detail.The structure and functional characteristics of each part of the laser microphone are analyzed structurally,and the working principle and process of the laser microphone are discussed.By studying the time-domain model of the optical fiber ring cavity,the relationship between the output power signal of the laser microphone and the length of the ring cavity is theoretically analyzed.Secondly,the characteristics of the voice and noise signals of the laser microphone are analyzed,and a method to achieve the time domain waveform alignment of the voice signal is proposed.Based on the characteristics of the distributed structure of the laser microphone and the wide range of sound pickup,the characteristics of the three kinds of additive noises of the voice signal and white noise,man-made noise and social noise recorded on the spot were analyzed.At the same time,in the method of aligning the time-domain waveform of the voice signal,by adding a square wave signal to the pure voice signal and using the synthesized signal as the input signal,after obtaining the output signal from the laser microphone,the input and output voice signals are intercepted from the marked nodes.This can achieve the alignment of two speech signals.This method is the optimization of the deep neural network speech enhancement algorithm,and the method is simple and easy,and the accuracy is high.Moreover,the noisy speech signal is enhanced based on the traditional speech enhancement algorithm.The classical wavelet transform algorithm has a good effect on non-stationary speech signal processing.In the experiment,the speech data training algorithm is used to select the number of decomposition layers,and the soft and hard compromise threshold function is used for experimental analysis.The Kalman filter method is the optimal estimate in the sense of minimum mean square error,which is suitable for the unstable signal of the laser microphone,and the optimal parameters for predicting the interference Q value and estimating the measurement variance R value are selected through experiments.Through experiments,it can be known that although the traditional speech enhancement algorithm has a small calculation amount and is simple and easy to implement,there will be a problem of residual noise.Finally,the noisy speech signal is enhanced based on the neural network structure.Neural network has strong adaptability and learning ability.In the deep neural network structure,the small batch gradient descent method is used to optimize the number of iterations,the network is trained with noisy speech signals,so that the network structure has anti-noise performance,and the BP algorithm is used to reverse optimization and update the network parameters to minimize Training time and data volume,and achieved better enhancement effect.Considering the limitation that the recurrent neural network cannot save information for a long time,here we combined the gated recurrent unit to enhance the speech signal.In the experiment,compared with the traditional speech enhancement algorithms,the advantages and disadvantages of each algorithm in the enhancement processing of noisy speech signals with different characteristics are analyzed through objective quantitative evaluation and spectrogram.